An advanced lightweight network with stepwise multiscale fusion in crowded scenes

Achieving a balance between high detection accuracy and lightweight network models poses a significant challenge in pedestrian detection. To address these issues, we introduce GSDC Net, an enhanced architecture inspired by the robust YOLOv7 model. Firstly, G-SDC module is designed by integrating smoothed dilated convolutions to bolster detection accuracy while simultaneously trimming the network’s computational footprint. And then, the SDCPPCSPC Module is advanced by employing Smoothed dilated convolutions to expand the receptive field, thereby capturing richer contextual information. To bolster feature fusion and augment the model’s capabilities, the bidirectional feature pyramid network and shuffle attention mechanisms are introduced. Through comparative experiments on the crowd human dataset, we have substantiated the effectiveness of our proposed model. The results indicate a notable improvement in mean average precision (mAP), reaching 83.88%. Additionally, the model’s size, and frames per second (FPS) have been optimized to 52.35 MB, and 82, respectively. These enhancements confirm the model’s exceptional balance among detection accuracy, inference speed, and model compactness.